We measure belief formation in a simple, large-scale experiment where participants, in different treatment conditions, are asked to provide forecasts of a stable random process. We use the data to analyze several expectations models put forward by the literature. Our findings are threefold. First, the rational expectations hypothesis is strongly rejected and we find little evidence of learning over time. Second, the data reject models where forecasters do not adjust the forecasting rule to properties of the data generating process. Third, an empirical model combining stickiness and forward-looking extrapolation does best in terms of explanatory power and stability. In this model, extrapolation quantitatively dominates, but stickiness makes past errors persist.